Bare Metal vs Managed GPU Cloud

Bare-metal infrastructure creates lower gross margins but exceptional customer stickiness. Customers lease dedicated hardware with no multi-tenancy, signing 3-5 year contracts for fixed capacity with guaranteed availability.

Margins typically run 25-35% gross. Customer CAC breaks even in 12-18 months, but churn is exceptionally low once deployed.

The capital intensity is brutal: you must purchase and deploy hardware before revenue arrives. The payoff is predictable revenue that sustains through business cycles. This model suits operators with patient capital, strong enterprise relationships, or hardware vendor support. Your customer base includes traditional enterprises migrating workloads, governments building sovereign capacity, and large research institutions.

Managed GPU services generate 45-65% gross margins by selling the same compute multiple times. Multiple customers share infrastructure through orchestration and per-second billing.

Churn is material (8-15% monthly), but the combination of high margins, low CAC, and operational leverage creates compelling unit economics for scaled operators. This model requires significant software investment: scheduling, orchestration, metering, isolation.

The upside is massive if you achieve density and utilization. Your customer base is fragmented (startups, researchers, small studios), creating both opportunity and complexity. Lambda Labs exemplified this developer-focused approach.

CoreWeave shifted from pure managed to hybrid (dedicated plus bursting capacity) after raising $200M+. At that scale, bare-metal capacity becomes economically viable: you can absorb capex lag and bet on enterprise contracts.

Lambda remained developer-focused. Their founding capital structure favored lower capex intensity, and they built deep integrations with the ML developer community.

Neither approach is universally superior. Both have thrived. Your choice comes down to three factors: capital access and risk tolerance, founder network and credibility with target customers, and willingness to build complex orchestration software for managed services.

The most successful operators (CoreWeave, Lambda, Crusoe Energy) now run both models simultaneously. Large customers get bare-metal commitments.

Smaller customers use bursting capacity on shared infrastructure. This requires operational discipline: separate teams, separate pricing, separate support models.

The payoff is capturing both margin profiles and maximising utilisation. The constraint is software complexity. Managing billing, isolation, and scheduling across dual models demands engineering investment that smaller operators cannot justify.

Bare-metal targets enterprises with long contracts, lower churn, 25-35% gross margins, and brutal capex requirements

Managed services target developers with 45-65% gross margins, higher churn (10-15% monthly), and significant software complexity

Your funding stage, customer relationships, and technical depth should determine which model fits; hybrid approaches work at scale

CoreWeave proved bare-metal viability at enterprise scale; Lambda proved developer-focused managed services can dominate mindshare

Success requires deep operational discipline; pricing, billing, support, and engineering differ fundamentally between models